Gemma
Gemma Loses 85% of Its Weights and Still Works: Specialized Models Are Cutting Local AI Costs
A developer pruned Gemma 4 26B-A4B to 15% of its size, retaining 92% agreement on one classification task while cutting memory use to 3.8GB.
1 article tagged with this topic
A developer pruned Gemma 4 26B-A4B to 15% of its size, retaining 92% agreement on one classification task while cutting memory use to 3.8GB.